What is the difference between SDXL base and refiner?

The base model does the bulk of the generation; the refiner is a second smaller model that polishes the final denoising steps for sharper detail. Most modern fine-tunes are trained to produce finished results in one pass, so the separate refiner is increasingly optional.

Do I still need the refiner with a fine-tune like Juggernaut XL?

Usually no. Community fine-tunes such as Juggernaut XL, RealVisXL, and DreamShaper XL already bake in the polish the refiner adds, so running them base-only is faster and often cleaner. Add the refiner only if fine edges or skin texture look soft.

Are SDXL 1.0 adapters compatible with these fine-tunes?

Yes. ControlNet, IP-Adapter, and LoRAs trained for SDXL 1.0 work across SDXL fine-tunes because they share the same architecture and latent space. Always match the adapter's SDXL version, not the specific checkpoint.

Does this tool generate images?

No. It is a static recommendation guide that runs in your browser. It points you to the right checkpoints and adapters to download or load in your own SDXL setup.

What is the SDXL Extension & Fine-tune Model Guide?

Interactive reference to the SDXL model ecosystem — base vs refiner roles, popular fine-tunes like RealVisXL, DreamShaper XL, and Juggernaut XL, compatible adapters such as ControlNet and IP-Adapter, and model picks by output style and use case. It runs free in your browser on Gera Tools, with nothing uploaded.

SDXL Extension & Fine-tune Model Guide

Name: SDXL Extension & Fine-tune Model Guide
Creator: Gera Tools
License: https://creativecommons.org/licenses/by/4.0/

SDXL extension & fine-tune model guide

The SDXL ecosystem is large: a base model, an optional refiner, dozens of community fine-tunes, and a stack of adapters like ControlNet, IP-Adapter, and LoRAs. Picking the right combination is the difference between fighting the model and getting clean results on the first batch. This guide maps your desired output style and use case to a concrete, compatible stack.

How it works

The tool holds a small knowledge base of the most widely used SDXL checkpoints and adapters and how they perform across styles. You select an output style and a use case, and it returns a recommended checkpoint, a note on whether the refiner is worth running, and the adapters that matter for that workflow. Because all SDXL fine-tunes share the base architecture, adapters are interchangeable as long as you match the SDXL version — so the guide focuses on which checkpoint sets the look and which adapters give you control.

The SDXL base and refiner roles

The original SDXL release shipped as a two-model pipeline. The base model (SDXL 1.0) does the main generation work: it denoise from pure noise to a rough but coherent image at latent resolution. The refiner is a smaller, separate model that takes the output of the base through the final denoising steps, polishing edges, recovering fine texture, and sharpening details that the base produces softly.

The two-stage pipeline produces good results but adds VRAM and generation time. The community responded by training fine-tunes designed to finish cleanly in a single pass without the refiner. Models like Juggernaut XL, RealVisXL, and DreamShaper XL incorporate the refinement capability into the checkpoint itself, which is why single-pass generation with a fine-tune typically equals or beats a base+refiner run — faster and with one fewer model to load.

Adapter compatibility across fine-tunes

All SDXL fine-tunes share the same underlying architecture and latent space as SDXL 1.0. This means adapters trained for SDXL 1.0 work across fine-tunes without modification. The adapters to know:

ControlNet XL — adds structural guidance from a depth map, edge detection, pose skeleton, or other conditioning image. Use it when you need to reproduce a specific composition, pose, or scene layout without prompt-guessing.

IP-Adapter XL — transfers the visual style, colour palette, or subject appearance from a reference image into the generated output. Particularly useful for character consistency across multiple generations or for matching a specific aesthetic reference.

LoRAs for SDXL — fine-tune slices trained on specific characters, artists, styles, or objects. Plug in with <lora:name:strength> syntax. LoRAs trained for SD 1.5 are not compatible — the architecture differs at a fundamental level.

Notes on the SDXL stack

Base + refiner is the original two-model pipeline. It still works, but most fine-tunes are tuned to finish in one pass, making the refiner optional.
RealVisXL leans photoreal, Juggernaut XL is a strong all-rounder, and DreamShaper XL balances realism with illustrative flexibility.
Adapters add control, not style. ControlNet locks pose and structure, IP-Adapter transfers a reference look, and LoRAs inject specific subjects.
Match the SDXL version when downloading adapters and LoRAs — 1.0-trained files work across 1.0 fine-tunes, but mismatched versions degrade output.
The base SDXL resolution is 1024×1024. Many fine-tunes support other native aspect ratios (1216×832 for widescreen, 832×1216 for portrait) — check the model card for the recommended resolution before generating.